There’s a moment in every PhD journey where your research stops being theoretical and starts keeping you up at night.

For me, it happened on a Tuesday. I was halfway through a stack of readings on AI ethics for my doctoral coursework — the kind of dense academic papers that make you feel simultaneously smarter and more confused — when I came across a psychological study that changed how I think about everything we build at Data Love.

The researchers, Maninger and Shank, ran a series of experiments where they presented people with identical moral violations. Same scenario, same harm, same outcome. The only difference? In one version, a human committed the violation. In the other, an AI system did. The result was striking and consistent: people blamed the AI less. Across nearly every moral category they tested, participants rated AI-caused harm as less wrong than human-caused harm. When it came to violations involving hierarchy and oppression, people were less likely to even classify what the AI did as a moral violation at all.

I set the paper down and stared at my screen. Because I realized this finding doesn’t just live in a lab. It lives in every food bank, every family assistance office, and every government agency that’s racing to deploy algorithmic tools without fully reckoning with what happens when those tools get it wrong.

What Is the Blame Gap — and Why Should You Care?

The “blame gap” is a term I use to describe the empirically demonstrated tendency for humans to assign less moral responsibility to AI systems than to people for equivalent harms. It’s not a matter of people being careless or indifferent. It’s deeper than that. Our entire cognitive architecture for moral judgment — how we assess blame, assign responsibility, and demand accountability — was built for interactions with other humans. When the agent causing harm is an algorithm, those mental circuits misfire.

This matters enormously for anyone working in human services, food security, housing, or public policy. Because if the public doesn’t hold AI systems accountable, then the institutions deploying those systems inherit a dangerous structural incentive: they can make consequential decisions that affect vulnerable populations, and when something goes wrong, the diffusion of responsibility across a “sociotechnical system” means nobody feels the full weight of blame.

Economists have a term for this: moral hazard. The consequences of bad decisions get spread across a system, while the benefits of efficiency get captured by specific groups. And the communities most affected — the families relying on SNAP benefits, the parents navigating childcare subsidies, the individuals seeking housing assistance — are the least positioned to push back.

This Isn’t Hypothetical. It’s Already Happened.

If the blame gap sounds abstract, consider two real-world cases that should alarm anyone who cares about equitable public services.

The Dutch Childcare Benefits Scandal. For years, the Netherlands’ Tax and Customs Administration used a self-learning algorithm to flag childcare benefit claims as potentially fraudulent. The system disproportionately targeted parents with non-Dutch nationality, treating dual citizenship as a risk factor. Over 35,000 families were falsely accused of fraud, ordered to repay tens of thousands of euros, and pushed into debt, poverty, and crisis. Parents lost homes. Relationships collapsed. At least one parent died by suicide. The Dutch government was forced to resign in 2021, and in 2022, the government publicly acknowledged that institutional racism was a root cause of the scandal.

What made it persist for so long? Algorithmic mediation diffused blame. When a human caseworker accuses a parent of fraud, the lines of accountability are clear. When an algorithm flags a claim and a bureaucratic process automatically issues a debt notice, the blame dissolves into the system itself. The EU Law Enforcement blog described it as a cautionary tale for any agency deploying algorithmic enforcement tools.

Australia’s Robodebt Scheme. Between 2016 and 2019, the Australian government’s automated debt recovery program used a flawed income-averaging algorithm to issue debt notices to public assistance recipients. The algorithm compared fortnightly Centrelink payments against averaged annual tax data — a method so fundamentally flawed that mathematicians later demonstrated it violated basic statistical principles. Over 450,000 Australians received debt notices, many of them incorrect. The Royal Commission that followed described the scheme’s underlying culture as one that viewed recipients as potential cheats. The program ultimately cost the government $565 million — more than it ever recovered — and was ruled unlawful.

Again, the pattern: the automation of the process reversed the burden of proof, reduced recipients’ access to human caseworkers, and made it extraordinarily difficult for vulnerable people to challenge decisions that were, in many cases, simply wrong. As the Blavatnik School of Government noted, the rush to introduce AI into public service delivery must be tempered by addressing legal and ethical issues first.

These aren’t edge cases. They’re the predictable consequences of deploying algorithmic systems in high-stakes environments without accounting for the blame gap.

The Double Failure: When Expertise and Public Pressure Both Break Down

Here’s what makes the blame gap so dangerous: it undermines both primary mechanisms of ethical accountability simultaneously.

The first mechanism is professional competence — the expectation that the people designing and deploying AI systems understand the ethical implications of what they’re building. But research by Khan, Akbar, and colleagues, surveying 99 AI practitioners and lawmakers across 20 countries, found that lack of ethical knowledge is the single most significant barrier to responsible AI adoption. More significant than absent legal frameworks. More significant than organizational resistance. The practitioners building these systems often don’t have the knowledge to recognize when they’re encoding harm.

The second mechanism is public pressure — the democratic expectation that citizens will hold institutions accountable when things go wrong. But Maninger and Shank’s findings show that the public lacks the psychological capacity to appropriately blame AI systems. When an algorithm informs a policy decision that reshapes food access for vulnerable communities, affected families are less likely to hold the institution accountable because, psychologically, “the data said so.”

Philosopher Iain Gabriel has argued that the AI alignment problem is fundamentally political rather than metaphysical — that it requires democratic consensus-building rather than the discovery of objective moral truth. I agree with that framing. But the blame gap means the democratic mechanisms we depend on to enforce alignment are themselves compromised. Even well-designed governance frameworks — transparency requirements, accountability structures, ethical review boards — risk becoming what some scholars call “managerial slogans” if they don’t account for the psychological reality that human moral intuition gives algorithms a pass.

If you’re interested in how this plays out in real organizations, I’ve explored some of these themes in past posts on Ethical AI in the Social Sector and AI for Good: How Mission-Driven Organizations Can Harness AI Responsibly.

Generative AI Makes It Worse

The emergence of large language models and generative AI compounds the blame gap in ways earlier research couldn’t have anticipated. Researchers applying anticipatory ethics to tools like ChatGPT have identified accountability as a high-impact ethical concern — one that the current discourse, focused on flashy issues like authorship and deepfakes, is largely missing at the systemic level.

Meanwhile, red-teaming research has shown that current AI safety mechanisms are brittle. Jailbreaking attacks can circumvent ethical guardrails, meaning we can’t rely on the AI system’s own alignment as a substitute for human accountability. When you combine that technical fragility with the blame gap, you arrive at a compounding problem: the systems most in need of external oversight are precisely those that human moral psychology is least equipped to hold accountable.

And there’s an environmental dimension that often goes undiscussed. Brute-force approaches to AI — where systems dump massive amounts of context into a language model and let it figure things out — are not just less reliable. They’re wasteful. Research benchmarking the environmental footprint of LLM inference has shown that the most energy-intensive models consume over 33 watt-hours per long prompt, more than 70 times the consumption of the most efficient systems. When you scale that to hundreds of millions of queries per day, the annual electricity consumption rivals that of tens of thousands of American homes. For organizations working in human services — where every dollar and every resource matters — wasteful AI architecture isn’t just a technical concern. It’s an ethical one.

At Data Love, this is something we take seriously. Our approach to AI-powered reporting is designed to minimize unnecessary computation through targeted retrieval and staged processing, rather than dumping everything into a model and hoping for the best. It’s not just about accuracy (though we’ve found that the architecture behind reporting matters enormously for getting numbers right). It’s about building systems that are environmentally responsible and aligned with the values of the communities we serve.

What This Means for Food Security and Human Services

These tensions aren’t abstract for me. At Data Love, my team builds AI-powered reporting pipelines for food security nonprofits and foundations, surfacing insights and recommendations that can inform food policy for state and local government agencies. Those policy decisions shape how, when, and where resources reach food-insecure families who rely on benefits programs and charitable food systems.

The blame gap operates across this entire chain. If an agency restructures distribution funding based on analytics, affected communities are psychologically less likely to attribute blame to decision-makers — because “the data said so.” Whose values shaped the metrics? Whose assumptions informed the recommendation? And here’s the knowledge-gap problem in action: policymakers and community stakeholders rarely have the technical literacy to interrogate algorithmic inputs.

No regulatory body oversees these algorithmic policy inputs. No maturity model exists for evaluating them. The communities most affected are least positioned to exercise the oversight that democratic accountability requires.

This is why we’ve been writing about the importance of being data-informed, not data-driven — because the framing matters. Data should support human judgment, not replace it. And it’s why using nonprofit data to influence public policy demands a level of responsibility and transparency that most current AI deployments simply don’t meet.

Toward an Accountability Architecture That Accounts for the Blame Gap

So what do we do? I don’t think the answer is to stop building AI tools for human services. The potential for well-designed systems to improve food access, housing stability, and health equity is too significant to walk away from. But I do think we need what I call blame-gap compensation: structural mechanisms that assign responsibility precisely because moral intuition won’t.

Here’s what that looks like in practice:

Human-in-the-loop governance as a genuine accountability anchor. Not rubber-stamping. Not a human clicking “approve” on outputs they don’t understand. Real, empowered human oversight at every decision point where algorithmic outputs inform policy that affects vulnerable populations. This is the lesson of both the Dutch scandal and Robodebt: when you remove meaningful human judgment from the chain, the system becomes capable of replicated harm at scale.

Closing the knowledge gap among affected communities. If policymakers and community stakeholders can’t interrogate algorithmic inputs, accountability is impossible. This means investing in data literacy programs, creating accessible documentation of how algorithmic recommendations are generated, and ensuring that communities have the tools and knowledge to ask hard questions about the systems shaping their access to resources. We’ve touched on this in our work exploring how nonprofit data can shape smarter policy decisions.

Anticipatory ethics before deployment, not after. Too often, ethical review happens after a system has been built and deployed — if it happens at all. Researchers in anticipatory technology ethics argue that we need to identify and address accountability risks before systems go live. For AI tools informing food policy, housing allocation, or benefits administration, this means conducting blame-gap assessments as part of the design process. Who will be affected? Who will be positioned to push back? And what structural accountability exists when — not if — the system gets it wrong?

Architectural choices that build in transparency. The way an AI system is built determines whether errors can even be found. Brute-force approaches — where a model receives a massive context dump and generates polished-looking outputs — are nearly impossible to audit. When a result is wrong, diagnosing the failure requires what I’ve heard called “prompt archaeology.” A staged, knowledge-first approach, by contrast, creates an execution trace at every step: what data was retrieved, what logic was applied, what assumptions were made, and what couldn’t be answered. That transparency isn’t just a technical feature. It’s an ethical requirement.

The Bottom Line

The blame gap is not a bug in human psychology that we can patch. It’s a feature of cognitive architecture built for human interaction. The organizational ethics challenge — for Data Love, for government agencies, for every institution deploying AI in high-stakes contexts — is to design governance that functions despite it.

Because when an algorithm informs a policy that reshapes food access for vulnerable communities, someone needs to be accountable. Regardless of whether moral intuition assigns blame.

---

This post was adapted from a thought paper written as part of my PhD research into Human-AI Collaboration in human services. If you care about ethical AI in the social sector, I’d love to have you along for the journey.

Subscribe to The Data Love Journal for field-tested insights on responsible AI, nonprofit data strategy, and the intersection of technology and social impact.

---

Sources & Further Reading

• Maninger, T., & Shank, D. B. (2022). Perceptions of violations by artificial and human actors across moral foundations. Computers in Human Behavior Reports.
• Gabriel, I. (2020). Artificial intelligence, values, and alignment. Minds and Machines.
• Khan, A. A., Akbar, M. A., et al. (2023). AI ethics: An empirical study on the views of practitioners and lawmakers. IEEE Transactions on Computational Social Systems.
• Stahl, B. C., & Eke, D. (2024). The ethics of ChatGPT. International Journal of Information Management.
• The Dutch benefits scandal: a cautionary tale for algorithmic enforcement — EU Law Enforcement
• Australia’s Robodebt scheme: A tragic case of public policy failure — Blavatnik School of Government
• The flawed algorithm at the heart of Robodebt — University of Melbourne
• Jegham, N., et al. (2025). How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference. arXiv.